Project: Predicting The Life time Value of Customers¶

Description: The Lifetime value of customers is given as the ratio of Cumulative Net Presnt Value and the total customers in the cohort. For our analysis, we shall use machine learning clustering algorith to create the cohorts and then use classification to classify the new customers into their respective cohorts and compute the LTV.¶

Importing Libraries¶

import numpy as np
import pandas as pd
import io
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.style as style
import matplotlib
import seaborn as sns
import datetime as dt
from IPython.display import Image
from datetime import datetime
from dateutil import relativedelta
import dill
from mpl_toolkits.mplot3d import Axes3D
import plotly.graph_objs as go
import plotly as py
import warnings
warnings.filterwarnings('ignore')
py.offline.init_notebook_mode(connected = True)

style.use('seaborn-darkgrid') #sets the size of the charts
style.use('ggplot')
matplotlib.rcParams['font.family'] = "serif"
sns.set_context('talk')

Importing Dataset¶

df = pd.read_csv("fact_bill_ltv_smry_data_0821_Cust.txt", sep = '|')

df.head()

Sample details of first customer in the Database:¶

df.columns = ['sno','cust_id', 'cust_estb_dt', 'cust_status_ind', 'recency', 'frequency', 'monetary', 'NewVZPlan_plan_cnt', 'VerizonUNL_plan_cnt', 'billed_acct_size', 'VoiceTextData_pplangrp_cnt', 'VoiceData_revtype_cnt', 'active_loan_cnt', 'birth_year', 'zip5_cd', 'Loand_equip_cnt', 'BeyondUnlmtd_plan_cnt', 'GoUnlimit_plan_cnt', 'IphoneDev_cnt', 'recurring_paychnl_cnt', 'CA_state_cd', 'Hispanic_ethnicity_cnt', 'data_ovrg_Yes_cnt', 'Edge_contract_cnt', 'CommunicStore_chnl_cnt', 'AndroidDev_cnt', 'WearablDev_cnt', 'SingleLine_plan_cnt', 'TX_state_cd', 'ATT_port_cnt', 'DataOnly_Revtype_cnt', 'MatureAdults_cnt', 'handset_paychnl_cnt', 'VoiceOnly_pplangrp_cnt', 'SngleCpleKids_cnt', 'FamTeens_cnt', 'Purch_equip_cnt', 'MTM_contract_cnt', 'T_Mobile_port_cnt', 'Sprint_port_cnt', 'LocalRetailer_chnl_cnt', 'TwoYear_contract_cnt', 'BasicDev_cnt', 'arpu_amt', 'gl_late_fees_amt', 'gl_tot_service_rev_amt', 'gl_recur_rev_amt', 'gl_nonrecur_rev_amt', 'pplan_access_acrtn', 'tot_revenue_acrtn', 'access_rev_acrtn', 'eqp_rev_acrtn', 'usage_rev_acrtn', 'data_overage_amt', 'voice_overage_amt']
df = df.drop('sno', axis=1)

df.head()

df.columns

Index(['cust_id', 'cust_estb_dt', 'cust_status_ind', 'recency', 'frequency',
       'monetary', 'NewVZPlan_plan_cnt', 'VerizonUNL_plan_cnt',
       'billed_acct_size', 'VoiceTextData_pplangrp_cnt',
       'VoiceData_revtype_cnt', 'active_loan_cnt', 'birth_year', 'zip5_cd',
       'Loand_equip_cnt', 'BeyondUnlmtd_plan_cnt', 'GoUnlimit_plan_cnt',
       'IphoneDev_cnt', 'recurring_paychnl_cnt', 'CA_state_cd',
       'Hispanic_ethnicity_cnt', 'data_ovrg_Yes_cnt', 'Edge_contract_cnt',
       'CommunicStore_chnl_cnt', 'AndroidDev_cnt', 'WearablDev_cnt',
       'SingleLine_plan_cnt', 'TX_state_cd', 'ATT_port_cnt',
       'DataOnly_Revtype_cnt', 'MatureAdults_cnt', 'handset_paychnl_cnt',
       'VoiceOnly_pplangrp_cnt', 'SngleCpleKids_cnt', 'FamTeens_cnt',
       'Purch_equip_cnt', 'MTM_contract_cnt', 'T_Mobile_port_cnt',
       'Sprint_port_cnt', 'LocalRetailer_chnl_cnt', 'TwoYear_contract_cnt',
       'BasicDev_cnt', 'arpu_amt', 'gl_late_fees_amt',
       'gl_tot_service_rev_amt', 'gl_recur_rev_amt', 'gl_nonrecur_rev_amt',
       'pplan_access_acrtn', 'tot_revenue_acrtn', 'access_rev_acrtn',
       'eqp_rev_acrtn', 'usage_rev_acrtn', 'data_overage_amt',
       'voice_overage_amt'],
      dtype='object')

len(df)

758196

df.isnull().sum()

cust_id                       0
cust_estb_dt                  0
cust_status_ind               0
recency                       0
frequency                     0
monetary                      0
NewVZPlan_plan_cnt            0
VerizonUNL_plan_cnt           0
billed_acct_size              0
VoiceTextData_pplangrp_cnt    0
VoiceData_revtype_cnt         0
active_loan_cnt               0
birth_year                    0
zip5_cd                       0
Loand_equip_cnt               0
BeyondUnlmtd_plan_cnt         0
GoUnlimit_plan_cnt            0
IphoneDev_cnt                 0
recurring_paychnl_cnt         0
CA_state_cd                   0
Hispanic_ethnicity_cnt        0
data_ovrg_Yes_cnt             0
Edge_contract_cnt             0
CommunicStore_chnl_cnt        0
AndroidDev_cnt                0
WearablDev_cnt                0
SingleLine_plan_cnt           0
TX_state_cd                   0
ATT_port_cnt                  0
DataOnly_Revtype_cnt          0
MatureAdults_cnt              0
handset_paychnl_cnt           0
VoiceOnly_pplangrp_cnt        0
SngleCpleKids_cnt             0
FamTeens_cnt                  0
Purch_equip_cnt               0
MTM_contract_cnt              0
T_Mobile_port_cnt             0
Sprint_port_cnt               0
LocalRetailer_chnl_cnt        0
TwoYear_contract_cnt          0
BasicDev_cnt                  0
arpu_amt                      0
gl_late_fees_amt              0
gl_tot_service_rev_amt        0
gl_recur_rev_amt              0
gl_nonrecur_rev_amt           0
pplan_access_acrtn            0
tot_revenue_acrtn             0
access_rev_acrtn              0
eqp_rev_acrtn                 0
usage_rev_acrtn               0
data_overage_amt              0
voice_overage_amt             0
dtype: int64

There are no Null values in the dataset.¶

df.head()

NOW = dt.datetime(2019,8,1) #Today's date (We take Aug 1, as we have data only until july 31)

Where are our customers located?¶

df_zip = pd.read_csv('us-zip-code-latitude-and-longitude.csv', sep=';')

Outlier Removal¶

df.describe()

# Outlier Removal:

def outlier_treatment(datacolumn):
 sorted(datacolumn)
 Q1,Q3 = np.percentile(datacolumn , [25,75])
 IQR = Q3 - Q1
 lower_range = Q1 - (1.5 * IQR)
 upper_range = Q3 + (1.5 * IQR)
 return lower_range,upper_range

print("Outlier boundaries for Monetary Value")
print(outlier_treatment(df['monetary']))

Outlier boundaries for Monetary Value
(-3335.15375, 9999.65625)

df = df[df.monetary > -3335.15375]
df = df[df.monetary < 9999.65625]

len(df)

735831

We now have 735,831 customers after outlier removal!

Removing unwanted columns:¶

df.columns

Index(['cust_id', 'cust_estb_dt', 'cust_status_ind', 'recency', 'frequency',
       'monetary', 'NewVZPlan_plan_cnt', 'VerizonUNL_plan_cnt',
       'billed_acct_size', 'VoiceTextData_pplangrp_cnt',
       'VoiceData_revtype_cnt', 'active_loan_cnt', 'birth_year', 'zip5_cd',
       'Loand_equip_cnt', 'BeyondUnlmtd_plan_cnt', 'GoUnlimit_plan_cnt',
       'IphoneDev_cnt', 'recurring_paychnl_cnt', 'CA_state_cd',
       'Hispanic_ethnicity_cnt', 'data_ovrg_Yes_cnt', 'Edge_contract_cnt',
       'CommunicStore_chnl_cnt', 'AndroidDev_cnt', 'WearablDev_cnt',
       'SingleLine_plan_cnt', 'TX_state_cd', 'ATT_port_cnt',
       'DataOnly_Revtype_cnt', 'MatureAdults_cnt', 'handset_paychnl_cnt',
       'VoiceOnly_pplangrp_cnt', 'SngleCpleKids_cnt', 'FamTeens_cnt',
       'Purch_equip_cnt', 'MTM_contract_cnt', 'T_Mobile_port_cnt',
       'Sprint_port_cnt', 'LocalRetailer_chnl_cnt', 'TwoYear_contract_cnt',
       'BasicDev_cnt', 'arpu_amt', 'gl_late_fees_amt',
       'gl_tot_service_rev_amt', 'gl_recur_rev_amt', 'gl_nonrecur_rev_amt',
       'pplan_access_acrtn', 'tot_revenue_acrtn', 'access_rev_acrtn',
       'eqp_rev_acrtn', 'usage_rev_acrtn', 'data_overage_amt',
       'voice_overage_amt'],
      dtype='object')

Customer Id seems to be the only column that can be removed as it has only unique values.

df = df.drop(['cust_id'], axis = 1)

df = df.drop(['cust_estb_dt'], axis = 1)

Pandas Profiling¶

import pandas_profiling
profile = pandas_profiling.ProfileReport(df)

profile

Correlation Check¶

f, ax = plt.subplots(figsize=(20, 12))
corr = df.corr()
hm = sns.heatmap(round(corr,2), annot=True, ax=ax, cmap="coolwarm",fmt='.2f',
            linewidths=.05)
f.subplots_adjust(top=0.93)
t= f.suptitle('Correlations between Features\n', fontsize=10)

corr_matrix = df.corr().abs()

#the matrix is symmetric so we need to extract upper triangle matrix without diagonal (k = 1)
sol = (corr_matrix.where(np.triu(np.ones(corr_matrix.shape), k=1).astype(np.bool))
                 .stack()
                 .sort_values(ascending=False))
#first element of sol series is the pair with the bigest correlation

sol

gl_tot_service_rev_amt      gl_recur_rev_amt              0.991984
monetary                    gl_tot_service_rev_amt        0.956043
billed_acct_size            VoiceTextData_pplangrp_cnt    0.948930
monetary                    gl_recur_rev_amt              0.948818
VoiceTextData_pplangrp_cnt  VoiceData_revtype_cnt         0.926589
billed_acct_size            VoiceData_revtype_cnt         0.925498
Loand_equip_cnt             Edge_contract_cnt             0.918416
recency                     frequency                     0.914661
SingleLine_plan_cnt         VoiceOnly_pplangrp_cnt        0.884635
cust_status_ind             recency                       0.879735
active_loan_cnt             Edge_contract_cnt             0.878532
                            Loand_equip_cnt               0.874786
billed_acct_size            active_loan_cnt               0.850371
VoiceTextData_pplangrp_cnt  active_loan_cnt               0.833709
VoiceData_revtype_cnt       active_loan_cnt               0.831770
cust_status_ind             frequency                     0.826603
VoiceData_revtype_cnt       Loand_equip_cnt               0.783133
VoiceTextData_pplangrp_cnt  Loand_equip_cnt               0.779297
billed_acct_size            Loand_equip_cnt               0.768819
                            Edge_contract_cnt             0.761489
VoiceData_revtype_cnt       Edge_contract_cnt             0.761278
VoiceTextData_pplangrp_cnt  Edge_contract_cnt             0.746970
monetary                    arpu_amt                      0.745987
billed_acct_size            arpu_amt                      0.733814
Edge_contract_cnt           arpu_amt                      0.732657
VoiceData_revtype_cnt       arpu_amt                      0.726980
Loand_equip_cnt             arpu_amt                      0.726792
VoiceTextData_pplangrp_cnt  arpu_amt                      0.721092
VoiceData_revtype_cnt       MTM_contract_cnt              0.709279
tot_revenue_acrtn           eqp_rev_acrtn                 0.708468
                                                            ...   
Purch_equip_cnt             voice_overage_amt             0.001793
SingleLine_plan_cnt         usage_rev_acrtn               0.001793
data_ovrg_Yes_cnt           FamTeens_cnt                  0.001732
VoiceOnly_pplangrp_cnt      voice_overage_amt             0.001714
BeyondUnlmtd_plan_cnt       usage_rev_acrtn               0.001648
ATT_port_cnt                voice_overage_amt             0.001633
DataOnly_Revtype_cnt        usage_rev_acrtn               0.001525
VoiceOnly_pplangrp_cnt      usage_rev_acrtn               0.001474
Hispanic_ethnicity_cnt      usage_rev_acrtn               0.001474
zip5_cd                     usage_rev_acrtn               0.001443
TX_state_cd                 usage_rev_acrtn               0.001388
DataOnly_Revtype_cnt        voice_overage_amt             0.001367
frequency                   T_Mobile_port_cnt             0.001243
MatureAdults_cnt            access_rev_acrtn              0.001203
Edge_contract_cnt           BasicDev_cnt                  0.001193
MTM_contract_cnt            usage_rev_acrtn               0.001051
Sprint_port_cnt             eqp_rev_acrtn                 0.000977
VoiceOnly_pplangrp_cnt      LocalRetailer_chnl_cnt        0.000932
birth_year                  pplan_access_acrtn            0.000877
Sprint_port_cnt             voice_overage_amt             0.000840
FamTeens_cnt                voice_overage_amt             0.000781
SingleLine_plan_cnt         voice_overage_amt             0.000734
birth_year                  CA_state_cd                   0.000668
handset_paychnl_cnt         eqp_rev_acrtn                 0.000628
DataOnly_Revtype_cnt        BasicDev_cnt                  0.000568
birth_year                  LocalRetailer_chnl_cnt        0.000383
TwoYear_contract_cnt        access_rev_acrtn              0.000367
birth_year                  usage_rev_acrtn               0.000303
gl_late_fees_amt            data_overage_amt              0.000235
TwoYear_contract_cnt        tot_revenue_acrtn             0.000228
Length: 1326, dtype: float64

We see that these are the highly correlated features, so we have to remove them:¶

gl_tot_service_rev_amt      gl_recur_rev_amt              0.991984
monetary                    gl_tot_service_rev_amt        0.956043
billed_acct_size            VoiceTextData_pplangrp_cnt    0.948930
monetary                    gl_recur_rev_amt              0.948818
VoiceTextData_pplangrp_cnt  VoiceData_revtype_cnt         0.926589
billed_acct_size            VoiceData_revtype_cnt         0.925498
Loand_equip_cnt             Edge_contract_cnt             0.918416

Let us remove:¶

gl_tot_service_rev_amt
VoiceTextData_pplangrp_cnt
gl_recur_rev_amt
VoiceData_revtype_cnt
Loand_equip_cnt

df = df.drop(['gl_tot_service_rev_amt',
    'VoiceTextData_pplangrp_cnt',
    'gl_recur_rev_amt',
    'VoiceData_revtype_cnt',
    'Loand_equip_cnt'], axis =1)

df.head()

Dropping Zip codes, as machine treats zip code with higher value as greater class:¶

df.zip5_cd.head()

0    12309
1    76104
2    12601
3    95337
4    14883
Name: zip5_cd, dtype: int64

df = df.drop('zip5_cd', axis = 1)

K Means Clustering:¶

X3 = df.iloc[: , :].values

X3

array([[ 1., 51., 25., ...,  0.,  0.,  0.],
       [ 1., 51., 21., ...,  0.,  0.,  0.],
       [ 1., 51., 22., ...,  0.,  0.,  0.],
       ...,
       [ 1., 51., 30., ...,  0.,  0.,  0.],
       [ 1., 51., 22., ...,  0.,  0.,  0.],
       [ 1., 51., 24., ...,  0.,  0.,  0.]])

Determining the Number of clusters using Elbow Curve:¶

# k means determine k
from sklearn.cluster import KMeans
from sklearn import metrics
from scipy.spatial.distance import cdist

distortions = []
K = range(1,10)
for k in K:
    kmeanModel = KMeans(n_clusters=k).fit(X3)
    kmeanModel.fit(X3)
    distortions.append(sum(np.min(cdist(X3, kmeanModel.cluster_centers_, 'euclidean'), axis=1)) / X3.shape[0])

# Plot the elbow
plt.plot(K, distortions, 'bx-')
plt.xlabel('k')
plt.ylabel('Distortion')
plt.title('The Elbow Method showing the optimal k')
plt.show()

The Elbow appears somewhere between 3 and 4. We shall choose 4 clusters to improve granularity and better segmentation.

from sklearn.cluster import KMeans
kmeans = KMeans(n_clusters=4, init='k-means++', random_state=0)
clusters = kmeans.fit_predict(X3)

df['Customer_Cluster'] = clusters

Checking the Attributes of Each Cluster:¶

df['Customer_Cluster'].value_counts()

0    287071
2    196863
3    160725
1     91172
Name: Customer_Cluster, dtype: int64

sns.countplot('Customer_Cluster',data=df)

<matplotlib.axes._subplots.AxesSubplot at 0x1dc0268b320>

fig = plt.figure(figsize=(8,6))
dx = fig.add_subplot(111, projection='3d')
colors = ['silver','midnightblue', 'maroon', 'gold']

for i in range(0,4):
    dx.scatter(df[df.Customer_Cluster == i].recency, 
               df[df.Customer_Cluster == i].frequency, 
               df[df.Customer_Cluster == i].monetary, 
               c = colors[i], 
               label = 'Cluster ' + str(i+1), 
               s=50)

dx.set_title('Clusters of clients')
dx.set_xlabel('Recency')
dx.set_ylabel('Frequency')
dx.set_zlabel('Monetary')
dx.legend()

<matplotlib.legend.Legend at 0x1dc024cd780>

It is difficult to visualize the clusters in 3 dimensions using just Recency, Frequency and Monetary values. Let us look at the Attributes of customers falling into each of these clusters:

Visualizing the Clusters in 3 Dimensions:¶

import plotly.graph_objs as go
import plotly as py

warnings.filterwarnings("ignore")
style.use('classic')
py.offline.init_notebook_mode(connected = True)
df = df.sample(frac=0.1, random_state=1)

trace1 = go.Scatter3d(
    x= df['recency'],
    y= df['frequency'],
    z= df['monetary'],
    mode='markers',
     marker=dict(
        color = df['Customer_Cluster'], 
        size= 20,
        line=dict(
            color= df['Customer_Cluster'],
            width= 12
        ),
        opacity=0.8
     )
)
data = [trace1]
layout = go.Layout(
#     margin=dict(
#         l=0,
#         r=0,
#         b=0,
#         t=0
#     )
    title= 'Customer_Cluster',
    scene = dict(
            xaxis = dict(title  = 'Recency'),
            yaxis = dict(title  = 'Frequency'),
            zaxis = dict(title  = 'Monetary Value')
        )
)
fig = go.Figure(data=data, layout=layout)
py.offline.iplot(fig)

Cluster 0:¶

pd.set_option('display.max_columns', 500)

df[df['Customer_Cluster']== 0].head(10)

Cluster 0:¶

df.head()

import pandas_profiling
profile1 = pandas_profiling.ProfileReport(df[df['Customer_Cluster']== 0])
profile1

Cluster 0 - Silver Findings:

1. Recency - Low (50 - 750)
2. Frequency - Low to Medium High (2 to 28)
3. Monetary - Low to High 
4. VoiceTextData_pplangrp_cnt - 1 to 5
5. Customer usually take Loans
6. IphoneDev_cnt - 67.6%
7. Total Revenue Acrtn - Negative
8. Some customers have voice overage
9. Customers mostly from TX
10. ARPU Amount - 128.36 (Mean)
11. billed_acct_size - 1.6 (Mean)
12. Birth Year - 1975 (Mean)
13. % of AT&T port in - 20.1%
14. % of Sprint_port in - 11.3%
15. % of T mobile port in - 12.5%
16. % of Android Devices - 42.2%
17. CA users - 49.3%
18. TX users - 26%
19. CommunicStore_chnl_cnt - 45.7%
20. Edge_contract_cnt - 85.8%
21. NewVZPlan_plan_cnt - 54.6%
22. SingleLine_plan_cnt - 2.4%
23. data_ovrg_Yes_cnt - 22%
24. tot_revenue_acrtn - -5.3379 (Mean)

Cluster 0:¶

df[df['Customer_Cluster']== 0].head(10)

Cluster 0 contains customers with medium low monetary values and are in the process of churning. We can name this cluster as "Silver"

Cluster 1:¶

profile2 = pandas_profiling.ProfileReport(df[df['Customer_Cluster']== 1])
profile2

Cluster 1 - Platinum Findings:

1. Recency - High (Mean - 58.759)
2. Frequency - High (Mean - 25.9)
3. Monetary - High (Mean - 7790)
6. IphoneDev_cnt - 87.2%
10. ARPU Amount - 305.87 (Mean)
11. billed_acct_size - 4.36 (Mean)
12. Birth Year - 1977 (Mean)
13. % of AT&T port in - 33.5%
14. % of Sprint_port in - 25.8%
15. % of T mobile port in - 22.7%
16. % of Android Devices - 58.8%
17. CA users - 46.8%
18. TX users - 35%
19. CommunicStore_chnl_cnt - 65.1%
20. Edge_contract_cnt - 99.2% (Mean - 3 lines per customer)
22. SingleLine_plan_cnt - 2.9%
23. data_ovrg_Yes_cnt - 11%
25. Hispanic Ethnicity - 38.1%
26. MTM_contract_cnt - 95.7%
27. Purch_equip_cnt - 43.5%
28. VerizonUNL_plan_cnt - 64.3%
29. WearablDev_cnt - 24.8%

df[df['Customer_Cluster']== 1].head(10)

Cluster 1 has customers making on-time payments and are loyal. We call this cluster as "Platinum"

Cluster 2:¶

profile3 = pandas_profiling.ProfileReport(df[df['Customer_Cluster']== 2])
profile3

Cluster 2 - Bronze Findings:

1. Recency - Very Low (Mean - 524.16)
2. Frequency - Very Low (Mean - 9.2962)
3. Monetary - Very Low (Mean - 702.76)
6. IphoneDev_cnt - 51.9%
10. ARPU Amount - 91.445 (Mean)
11. billed_acct_size - 1.3875 (Mean)
12. Birth Year - 1971 (Mean)
13. % of AT&T port in - 13.1%
14. % of Sprint_port in - 6.5%
15. % of T mobile port in - 12.2%
16. % of Android Devices - 34%
17. CA users - 49.1%
18. TX users - 27.1%
19. CommunicStore_chnl_cnt - 40.2%
20. Edge_contract_cnt - 44.2% 
22. SingleLine_plan_cnt - 13.2%
23. data_ovrg_Yes_cnt - 13.4%
25. Hispanic Ethnicity - 26.7%
27. Purch_equip_cnt - 43.5%
29. WearablDev_cnt - 3.2%
30. active_loan_cnt - 50%
31. Total Revenue Acrtn - -30.913(Mean)

df[df['Customer_Cluster']== 2].head(10)

Cluster 2 contains customers who are either churned or are having trouble paying their monthly phone bills. We can name this cluster as "bronze"

Cluster 3:¶

profile4 = pandas_profiling.ProfileReport(df[df['Customer_Cluster']== 3])
profile4

Cluster 3 - Gold Findings:

1. Recency - High (Mean - 70.6)
2. Frequency - High (Mean - 24.9)
3. Monetary - High (Mean - 5003.7)
6. IphoneDev_cnt - 80%
10. ARPU Amount - 208.06 (Mean)
11. billed_acct_size - 2.8353 (Mean)
12. Birth Year - 1977 (Mean)
13. % of AT&T port in - 29.1%
14. % of Sprint_port in - 18%
15. % of T mobile port in - 17%
16. % of Android Devices - 49%
17. CA users - 47.7%
18. TX users - 32%
19. CommunicStore_chnl_cnt - 58%
20. Edge_contract_cnt - 97.4% 
22. SingleLine_plan_cnt - 3%
23. data_ovrg_Yes_cnt - 21%
27. Purch_equip_cnt - 43.5%
29. WearablDev_cnt - 16%
30. active_loan_cnt - 50%
31. Total Revenue Acrtn - -30.913(Mean)
32. MTM_contract_cnt - 90%

df[df['Customer_Cluster']== 3].head(10)

Cluster 3 contains average customers with medium monetary values. We can name this cluster as "Gold"

df.columns

Index(['cust_status_ind', 'recency', 'frequency', 'monetary',
       'NewVZPlan_plan_cnt', 'VerizonUNL_plan_cnt', 'billed_acct_size',
       'active_loan_cnt', 'birth_year', 'BeyondUnlmtd_plan_cnt',
       'GoUnlimit_plan_cnt', 'IphoneDev_cnt', 'recurring_paychnl_cnt',
       'CA_state_cd', 'Hispanic_ethnicity_cnt', 'data_ovrg_Yes_cnt',
       'Edge_contract_cnt', 'CommunicStore_chnl_cnt', 'AndroidDev_cnt',
       'WearablDev_cnt', 'SingleLine_plan_cnt', 'TX_state_cd', 'ATT_port_cnt',
       'DataOnly_Revtype_cnt', 'MatureAdults_cnt', 'handset_paychnl_cnt',
       'VoiceOnly_pplangrp_cnt', 'SngleCpleKids_cnt', 'FamTeens_cnt',
       'Purch_equip_cnt', 'MTM_contract_cnt', 'T_Mobile_port_cnt',
       'Sprint_port_cnt', 'LocalRetailer_chnl_cnt', 'TwoYear_contract_cnt',
       'BasicDev_cnt', 'arpu_amt', 'gl_late_fees_amt', 'gl_nonrecur_rev_amt',
       'pplan_access_acrtn', 'tot_revenue_acrtn', 'access_rev_acrtn',
       'eqp_rev_acrtn', 'usage_rev_acrtn', 'data_overage_amt',
       'voice_overage_amt', 'Customer_Cluster'],
      dtype='object')

Cluster Naming:¶

1. Cluster 0 : Silver 
2. Cluster 1 : Platinum
3. Cluster 2 : Bronze
4. Cluster 3 : Gold

# Renaming the clusters in the original dataframe:
df['Customer_Cluster'] = np.select([df.Customer_Cluster == 0, df.Customer_Cluster == 1, df.Customer_Cluster == 2, df.Customer_Cluster == 3], 
                        ['Silver','Platinum','Bronze','Gold'])

df.Customer_Cluster.value_counts()

Silver      28578
Bronze      19777
Gold        16041
Platinum     9187
Name: Customer_Cluster, dtype: int64

sns.countplot('Customer_Cluster',data=df)

<matplotlib.axes._subplots.AxesSubplot at 0x1dc1963da90>

df1 = pd.read_csv("fact_bill_ltv_smry_data_0821_Cust.txt", sep = '|')

df1.head()

df1.columns = ['sno','cust_id', 'cust_estb_dt', 'cust_status_ind', 'recency', 'frequency', 'monetary', 'NewVZPlan_plan_cnt', 'VerizonUNL_plan_cnt', 'billed_acct_size', 'VoiceTextData_pplangrp_cnt', 'VoiceData_revtype_cnt', 'active_loan_cnt', 'birth_year', 'zip5_cd', 'Loand_equip_cnt', 'BeyondUnlmtd_plan_cnt', 'GoUnlimit_plan_cnt', 'IphoneDev_cnt', 'recurring_paychnl_cnt', 'CA_state_cd', 'Hispanic_ethnicity_cnt', 'data_ovrg_Yes_cnt', 'Edge_contract_cnt', 'CommunicStore_chnl_cnt', 'AndroidDev_cnt', 'WearablDev_cnt', 'SingleLine_plan_cnt', 'TX_state_cd', 'ATT_port_cnt', 'DataOnly_Revtype_cnt', 'MatureAdults_cnt', 'handset_paychnl_cnt', 'VoiceOnly_pplangrp_cnt', 'SngleCpleKids_cnt', 'FamTeens_cnt', 'Purch_equip_cnt', 'MTM_contract_cnt', 'T_Mobile_port_cnt', 'Sprint_port_cnt', 'LocalRetailer_chnl_cnt', 'TwoYear_contract_cnt', 'BasicDev_cnt', 'arpu_amt', 'gl_late_fees_amt', 'gl_tot_service_rev_amt', 'gl_recur_rev_amt', 'gl_nonrecur_rev_amt', 'pplan_access_acrtn', 'tot_revenue_acrtn', 'access_rev_acrtn', 'eqp_rev_acrtn', 'usage_rev_acrtn', 'data_overage_amt', 'voice_overage_amt']
df1 = df1.drop('sno', axis=1)

df1 = df1[df1.monetary > -3335.15375]
df1 = df1[df1.monetary < 9999.65625]

df2 = df['Customer_Cluster']

result = pd.concat([df1, df2], axis=1)

result.head()

#result.to_csv (r'Final_table.csv', index = None, header=True)

Classification Models¶

df.head()

df.info()

<class 'pandas.core.frame.DataFrame'>
Int64Index: 73583 entries, 56841 to 491409
Data columns (total 47 columns):
cust_status_ind           73583 non-null int64
recency                   73583 non-null int64
frequency                 73583 non-null int64
monetary                  73583 non-null float64
NewVZPlan_plan_cnt        73583 non-null int64
VerizonUNL_plan_cnt       73583 non-null int64
billed_acct_size          73583 non-null int64
active_loan_cnt           73583 non-null int64
birth_year                73583 non-null int64
BeyondUnlmtd_plan_cnt     73583 non-null int64
GoUnlimit_plan_cnt        73583 non-null int64
IphoneDev_cnt             73583 non-null int64
recurring_paychnl_cnt     73583 non-null int64
CA_state_cd               73583 non-null int64
Hispanic_ethnicity_cnt    73583 non-null int64
data_ovrg_Yes_cnt         73583 non-null int64
Edge_contract_cnt         73583 non-null int64
CommunicStore_chnl_cnt    73583 non-null int64
AndroidDev_cnt            73583 non-null int64
WearablDev_cnt            73583 non-null int64
SingleLine_plan_cnt       73583 non-null int64
TX_state_cd               73583 non-null int64
ATT_port_cnt              73583 non-null int64
DataOnly_Revtype_cnt      73583 non-null int64
MatureAdults_cnt          73583 non-null int64
handset_paychnl_cnt       73583 non-null int64
VoiceOnly_pplangrp_cnt    73583 non-null int64
SngleCpleKids_cnt         73583 non-null int64
FamTeens_cnt              73583 non-null int64
Purch_equip_cnt           73583 non-null int64
MTM_contract_cnt          73583 non-null int64
T_Mobile_port_cnt         73583 non-null int64
Sprint_port_cnt           73583 non-null int64
LocalRetailer_chnl_cnt    73583 non-null int64
TwoYear_contract_cnt      73583 non-null int64
BasicDev_cnt              73583 non-null int64
arpu_amt                  73583 non-null float64
gl_late_fees_amt          73583 non-null float64
gl_nonrecur_rev_amt       73583 non-null float64
pplan_access_acrtn        73583 non-null float64
tot_revenue_acrtn         73583 non-null float64
access_rev_acrtn          73583 non-null float64
eqp_rev_acrtn             73583 non-null float64
usage_rev_acrtn           73583 non-null float64
data_overage_amt          73583 non-null float64
voice_overage_amt         73583 non-null float64
Customer_Cluster          73583 non-null object
dtypes: float64(11), int64(35), object(1)
memory usage: 29.4+ MB

Distribution of Output Variable:¶

plt.figure(figsize=(20,7))
plt.subplot(121)
df["Customer_Cluster"].value_counts().plot.pie(autopct = "%1.0f%%",colors = sns.color_palette("prism",4),startangle = 60,labels=["Silver","Bronze","Gold","Platinum"],
                                                                        wedgeprops={"linewidth":2,"edgecolor":"k"},explode=[0,0.1,0.2,0.3],shadow =True)
plt.title("Distribution of Target Variable")

plt.subplot(122)
ax = df["Customer_Cluster"].value_counts().plot(kind="barh")

for i,j in enumerate(df["Customer_Cluster"].value_counts().values):
    ax.text(.7,i,j,weight = "bold",fontsize=20)

plt.title("Count of Cluster")
plt.show()

df.columns

Index(['cust_status_ind', 'recency', 'frequency', 'monetary',
       'NewVZPlan_plan_cnt', 'VerizonUNL_plan_cnt', 'billed_acct_size',
       'active_loan_cnt', 'birth_year', 'BeyondUnlmtd_plan_cnt',
       'GoUnlimit_plan_cnt', 'IphoneDev_cnt', 'recurring_paychnl_cnt',
       'CA_state_cd', 'Hispanic_ethnicity_cnt', 'data_ovrg_Yes_cnt',
       'Edge_contract_cnt', 'CommunicStore_chnl_cnt', 'AndroidDev_cnt',
       'WearablDev_cnt', 'SingleLine_plan_cnt', 'TX_state_cd', 'ATT_port_cnt',
       'DataOnly_Revtype_cnt', 'MatureAdults_cnt', 'handset_paychnl_cnt',
       'VoiceOnly_pplangrp_cnt', 'SngleCpleKids_cnt', 'FamTeens_cnt',
       'Purch_equip_cnt', 'MTM_contract_cnt', 'T_Mobile_port_cnt',
       'Sprint_port_cnt', 'LocalRetailer_chnl_cnt', 'TwoYear_contract_cnt',
       'BasicDev_cnt', 'arpu_amt', 'gl_late_fees_amt', 'gl_nonrecur_rev_amt',
       'pplan_access_acrtn', 'tot_revenue_acrtn', 'access_rev_acrtn',
       'eqp_rev_acrtn', 'usage_rev_acrtn', 'data_overage_amt',
       'voice_overage_amt', 'Customer_Cluster'],
      dtype='object')

Since we have already removed highly correlated features before performing Clustering, there is no need to apply correlation check again.

Classification for 2019 Customers (New customers):¶

from sklearn.model_selection import train_test_split

X = df[['NewVZPlan_plan_cnt', 'VerizonUNL_plan_cnt', 'billed_acct_size',
       'active_loan_cnt', 'birth_year', 'BeyondUnlmtd_plan_cnt',
       'GoUnlimit_plan_cnt', 'IphoneDev_cnt', 'recurring_paychnl_cnt',
       'CA_state_cd', 'Hispanic_ethnicity_cnt', 'data_ovrg_Yes_cnt',
       'Edge_contract_cnt', 'CommunicStore_chnl_cnt', 'AndroidDev_cnt',
       'WearablDev_cnt', 'SingleLine_plan_cnt', 'TX_state_cd', 'ATT_port_cnt',
       'DataOnly_Revtype_cnt', 'MatureAdults_cnt', 'handset_paychnl_cnt',
       'VoiceOnly_pplangrp_cnt', 'SngleCpleKids_cnt', 'FamTeens_cnt',
       'Purch_equip_cnt', 'MTM_contract_cnt', 'T_Mobile_port_cnt',
       'Sprint_port_cnt', 'LocalRetailer_chnl_cnt', 'TwoYear_contract_cnt',
       'BasicDev_cnt', 'arpu_amt','data_overage_amt','voice_overage_amt']]
y = df['Customer_Cluster']

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

X_train.head()

Standard Scaler for Feature Scaling¶

# Feature Scaling:

from sklearn.preprocessing import StandardScaler

sc = StandardScaler()
X_train = sc.fit_transform(X_train)
X_test = sc.transform(X_test)

Decision Tree Classifier¶

from sklearn.metrics import roc_curve, precision_recall_curve, auc, make_scorer, recall_score, accuracy_score, classification_report, precision_score, confusion_matrix

# ---- Decision Tree -----------
from sklearn import tree


dtree = tree.DecisionTreeClassifier(criterion='entropy', max_depth=5)
dtree.fit(X_train, y_train)
y_pred_dtree = dtree.predict(X_test) 
print('Accuracy of decision tree on training set: {:.2f}'
     .format(dtree.score(X_train, y_train)))
print('Accuracy of decision tree  on validation set: {:.2f}'
     .format(dtree.score(X_test, y_test)))
print('Clasification report of Decision Treeclassifier on validation set: \n{}'
     .format(classification_report(y_test, y_pred_dtree)))
print(confusion_matrix(y_test, y_pred_dtree))

Accuracy of decision tree on training set: 0.74
Accuracy of decision tree  on validation set: 0.74
Clasification report of Decision Treeclassifier on validation set: 
              precision    recall  f1-score   support

      Bronze       0.81      0.69      0.75      3931
        Gold       0.66      0.71      0.69      3243
    Platinum       0.69      0.81      0.75      1828
      Silver       0.76      0.76      0.76      5715

   micro avg       0.74      0.74      0.74     14717
   macro avg       0.73      0.74      0.74     14717
weighted avg       0.74      0.74      0.74     14717

[[2723  217  161  830]
 [  12 2299  356  576]
 [   0  341 1487    0]
 [ 616  609  138 4352]]

Random Forest Classifier¶

# ---- Random Forest -----------
from sklearn.ensemble import RandomForestClassifier

rfc = RandomForestClassifier(n_estimators=100)
rfc.fit(X_train, y_train)
y_pred_rfc = rfc.predict(X_test) 

print('Accuracy of Random Forest on training set: {:.2f}'
     .format(rfc.score(X_train, y_train)))
print('Accuracy of Random Forest  on validation set: {:.2f}'
     .format(rfc.score(X_test, y_test)))
print('Clasification report of Random Forest on validation set: \n{}'
     .format(classification_report(y_test, y_pred_rfc)))
print(confusion_matrix(y_test, y_pred_rfc))

Accuracy of Random Forest on training set: 1.00
Accuracy of Random Forest  on validation set: 0.79
Clasification report of Random Forest on validation set: 
              precision    recall  f1-score   support

      Bronze       0.86      0.76      0.81      3931
        Gold       0.70      0.77      0.73      3243
    Platinum       0.77      0.80      0.79      1828
      Silver       0.80      0.81      0.80      5715

   micro avg       0.79      0.79      0.79     14717
   macro avg       0.78      0.78      0.78     14717
weighted avg       0.79      0.79      0.79     14717

[[2977  145   51  758]
 [  43 2481  292  427]
 [  16  340 1468    4]
 [ 428  563   95 4629]]

Finding the best parameters for Random forest:¶

# Number of features to consider at every split
max_features = ['auto', 'sqrt']
# Maximum number of levels in tree
max_depth = [int(x) for x in np.linspace(10, 80, num = 11)]
max_depth.append(None)
# Minimum number of samples required to split a node
min_samples_split = [2, 5, 10]
# Minimum number of samples required at each leaf node
min_samples_leaf = [1, 2, 4]
# Method of selecting samples for training each tree
bootstrap = [True, False]

random_grid = {
               'max_features': max_features,
               'max_depth': max_depth,
               'min_samples_split': min_samples_split,
               'min_samples_leaf': min_samples_leaf,
               'bootstrap': bootstrap}
print(random_grid)

{'max_features': ['auto', 'sqrt'], 'max_depth': [10, 17, 24, 31, 38, 45, 52, 59, 66, 73, 80, None], 'min_samples_split': [2, 5, 10], 'min_samples_leaf': [1, 2, 4], 'bootstrap': [True, False]}

from sklearn.model_selection import RandomizedSearchCV
# Use the random grid to search for best hyperparameters
# First create the base model to tune
rf = RandomForestClassifier()
# Random search of parameters, using 3 fold cross validation, 
# search across 100 different combinations, and use all available cores
rf_random = RandomizedSearchCV(estimator = rf, param_distributions = random_grid, n_iter = 10, cv = 3, verbose=2, random_state=42, n_jobs = -1)
# Fit the random search model
rf_random.fit(X_train, y_train)
y_pred_rf = rf_random.predict(X_test)

Fitting 3 folds for each of 10 candidates, totalling 30 fits

[Parallel(n_jobs=-1)]: Using backend LokyBackend with 4 concurrent workers.
[Parallel(n_jobs=-1)]: Done  30 out of  30 | elapsed:   22.2s finished

rf_random.best_params_

{'min_samples_split': 10,
 'min_samples_leaf': 2,
 'max_features': 'sqrt',
 'max_depth': 73,
 'bootstrap': False}

Random Forest tuned for Hyper Parameters¶

# ---- Random Forest with best parameters -----------
from sklearn.ensemble import RandomForestClassifier

rfc = RandomForestClassifier(n_estimators=10,min_samples_split= 10,min_samples_leaf= 2,max_features='sqrt',max_depth=73,bootstrap=False)
rfc.fit(X_train, y_train)
y_pred_rfc = rfc.predict(X_test)

print('Accuracy of Random Forest on training set: {:.2f}'
     .format(rfc.score(X_train, y_train)))
print('Accuracy of Random Forest  on validation set: {:.2f}'
     .format(rfc.score(X_test, y_test)))
print('Classification report of Random Forest on validation set: \n{}'
     .format(classification_report(y_test, y_pred_rfc)))
print(confusion_matrix(y_test, y_pred_rfc))

Accuracy of Random Forest on training set: 0.94
Accuracy of Random Forest  on validation set: 0.78
Classification report of Random Forest on validation set: 
              precision    recall  f1-score   support

      Bronze       0.85      0.76      0.80      3931
        Gold       0.70      0.75      0.73      3243
    Platinum       0.77      0.79      0.78      1828
      Silver       0.79      0.81      0.80      5715

   micro avg       0.78      0.78      0.78     14717
   macro avg       0.78      0.78      0.78     14717
weighted avg       0.78      0.78      0.78     14717

[[2977  138   51  765]
 [  45 2445  291  462]
 [  18  355 1446    9]
 [ 458  545  102 4610]]

XG Boost Classifier¶

#XGBoost Multiclassification Model
from xgboost import XGBClassifier

xgb = XGBClassifier()
xgb.fit(X_train, y_train)
y_pred_xgb = xgb.predict(X_test) 

print('Accuracy of XGBoost on training set: {:.2f}'
     .format(xgb.score(X_train, y_train)))
print('Accuracy of XGBoost  on validation set: {:.2f}'
     .format(xgb.score(X_test, y_test)))
print('Classification report of XGBoost on validation set: \n{}'
     .format(classification_report(y_test, y_pred_xgb)))
print(confusion_matrix(y_test, y_pred_xgb))

Accuracy of XGBoost on training set: 0.78
Accuracy of XGBoost  on validation set: 0.78
Classification report of XGBoost on validation set: 
              precision    recall  f1-score   support

      Bronze       0.86      0.71      0.78      3931
        Gold       0.71      0.76      0.73      3243
    Platinum       0.76      0.82      0.79      1828
      Silver       0.78      0.82      0.80      5715

   micro avg       0.78      0.78      0.78     14717
   macro avg       0.78      0.78      0.78     14717
weighted avg       0.78      0.78      0.78     14717

[[2804  169   80  878]
 [  40 2459  283  461]
 [  19  306 1502    1]
 [ 397  526  104 4688]]

Naive Bayes Classifier¶

# ---- Naive Bayes -----------
from sklearn.naive_bayes import GaussianNB

gnb = GaussianNB()
gnb.fit(X_train, y_train)
y_pred_gnb = gnb.predict(X_test) 
print('Accuracy of GNB classifier on training set: {:.2f}'
     .format(gnb.score(X_train, y_train)))
print('Accuracy of GNB classifier on validation set: {:.2f}'
     .format(gnb.score(X_test, y_test)))
print('Classification report of Gaussian Naive Bayes on validation set: \n{}'
     .format(classification_report(y_test, y_pred_gnb)))
print(confusion_matrix(y_test, y_pred_gnb))

Accuracy of GNB classifier on training set: 0.58
Accuracy of GNB classifier on validation set: 0.58
Classification report of Gaussian Naive Bayes on validation set: 
              precision    recall  f1-score   support

      Bronze       0.84      0.36      0.50      3931
        Gold       0.48      0.40      0.44      3243
    Platinum       0.57      0.61      0.59      1828
      Silver       0.57      0.84      0.68      5715

   micro avg       0.58      0.58      0.58     14717
   macro avg       0.61      0.55      0.55     14717
weighted avg       0.62      0.58      0.57     14717

[[1405  284  154 2088]
 [  39 1294  500 1410]
 [   8  594 1115  111]
 [ 228  522  184 4781]]

Ada Boost Classifier¶

# ---- Ada Boost -----------
from sklearn.ensemble import AdaBoostClassifier
abc = AdaBoostClassifier(random_state=1)
abc.fit(X_train, y_train)
y_pred_abc = abc.predict(X_test) 

print('Accuracy of AdaBoost on training set: {:.2f}'
     .format(abc.score(X_train, y_train)))
print('Accuracy of AdaBoost  on validation set: {:.2f}'
     .format(abc.score(X_test, y_test)))
print('Classification report of Ada Boost on validation set: \n{}'
     .format(classification_report(y_test, y_pred_abc)))
print(confusion_matrix(y_test, y_pred_abc))

Accuracy of AdaBoost on training set: 0.70
Accuracy of AdaBoost  on validation set: 0.69
Classification report of Ada Boost on validation set: 
              precision    recall  f1-score   support

      Bronze       0.76      0.52      0.62      3931
        Gold       0.66      0.68      0.67      3243
    Platinum       0.72      0.82      0.77      1828
      Silver       0.67      0.78      0.72      5715

   micro avg       0.69      0.69      0.69     14717
   macro avg       0.70      0.70      0.69     14717
weighted avg       0.70      0.69      0.69     14717

[[2027  147   64 1693]
 [ 115 2202  425  501]
 [  27  284 1508    9]
 [ 487  696  102 4430]]

Confusion Matrix for the Best Model¶

# Creates a confusion matrix
cm = confusion_matrix(y_test, y_pred_rfc) 

# Transform to df for easier plotting
cm_df = pd.DataFrame(cm,
                     index = ['Platinum','Gold','Silver','Bronze'], 
                     columns = ['Platinum','Gold','Silver','Bronze'])

plt.figure(figsize=(8,6))
sns.heatmap(cm_df, annot=True,fmt='2.0f')
plt.title('Random Forest \nAccuracy:{0:.3f}'.format(accuracy_score(y_test, y_pred_rfc)))
plt.ylabel('True label')
plt.xlabel('Predicted label')
plt.show()

Shap Feature Importance: (Model Explainability)¶

We now try to obtain the features that contribute the most to our Random forest model. We can use Shap feature explanation for the same.

import shap
pd.set_option("display.max_columns", 96)
pd.set_option("display.max_rows", 96)

plt.rcParams['figure.figsize'] = (12, 9)
plt.style.use('ggplot')

shap.initjs()

# DF, based on which importance is checked
X_importance = X_test

# Explain model predictions using shap library:
explainer = shap.TreeExplainer(rfc)
shap_values = explainer.shap_values(X_importance)

# Plot summary_plot
shap.summary_plot(shap_values, X)

Improving the Model Accuracy: (Grid Search)¶

grid_param = {
    'n_estimators': [100, 200, 300],
    'criterion': ['gini', 'entropy'],
    'bootstrap': [True, False]
}

from sklearn.model_selection import GridSearchCV

gd_sr = GridSearchCV(estimator=rfc,
                     param_grid=grid_param,
                     scoring='accuracy',
                     cv=5,
                     n_jobs=-1)

gd_sr.fit(X_train, y_train)

GridSearchCV(cv=5, error_score='raise-deprecating',
       estimator=RandomForestClassifier(bootstrap=False, class_weight=None, criterion='gini',
            max_depth=73, max_features='sqrt', max_leaf_nodes=None,
            min_impurity_decrease=0.0, min_impurity_split=None,
            min_samples_leaf=2, min_samples_split=10,
            min_weight_fraction_leaf=0.0, n_estimators=10, n_jobs=None,
            oob_score=False, random_state=None, verbose=0,
            warm_start=False),
       fit_params=None, iid='warn', n_jobs=-1,
       param_grid={'n_estimators': [100, 200, 300], 'criterion': ['gini', 'entropy'], 'bootstrap': [True, False]},
       pre_dispatch='2*n_jobs', refit=True, return_train_score='warn',
       scoring='accuracy', verbose=0)

The Best Estimators Obtained for the model after Grid Search CV are:¶

best_parameters = gd_sr.best_params_
print(best_parameters)

{'bootstrap': False, 'criterion': 'gini', 'n_estimators': 300}

Results of Best Estimator:¶

best_result = gd_sr.best_score_
print(best_result)

0.7871946454659736

The accuracy we have obtained after tuning the best model is 78.71%. We can further improve the performance of this model by stacking. We can combaine Random Forest and XGBoost algorithms to make better predictions!

Stacking¶

from mlxtend.classifier import StackingClassifier
from mlxtend.plotting import plot_learning_curves
from mlxtend.plotting import plot_decision_regions
from sklearn import model_selection

clf1 = RandomForestClassifier(random_state=0,n_estimators=300, max_depth=73, criterion = 'gini')
clf2 = XGBClassifier()
clf3 = GaussianNB()
rfc = RandomForestClassifier(n_estimators=100)
sclf = StackingClassifier(classifiers=[clf1, clf2, clf3], 
                          meta_classifier=rfc)

# Import label encoder 
from sklearn import preprocessing 
  
# label_encoder object knows how to understand word labels. 
label_encoder = preprocessing.LabelEncoder() 
  
# Encode labels in column 'species'. 
y_train = label_encoder.fit_transform(y_train) 
y_test = label_encoder.fit_transform(y_test)

sclf = StackingClassifier(classifiers=[clf1, clf2, clf3], 
                          meta_classifier=rfc) 
sclf.fit(X_train, y_train)
y_pred_sclf = sclf.predict(X_test) 

print('Accuracy of Stacking classifier on training set: {:.2f}'
     .format(sclf.score(X_train, y_train)))
print('Accuracy of Stacked Classifier on test set: {:.2f}'
     .format(sclf.score(X_test, y_test)))
print('Clasification report of Stacked Classifier on Test set: \n{}'
     .format(classification_report(y_test, y_pred_sclf)))
print(confusion_matrix(y_test, y_pred_sclf))

Accuracy of Stacking classifier on training set: 1.00
Accuracy of Stacked Classifier on test set: 0.79
Clasification report of Stacked Classifier on Test set: 
              precision    recall  f1-score   support

           0       0.86      0.76      0.80      3931
           1       0.70      0.77      0.74      3243
           2       0.77      0.80      0.79      1828
           3       0.80      0.81      0.80      5715

   micro avg       0.79      0.79      0.79     14717
   macro avg       0.78      0.78      0.78     14717
weighted avg       0.79      0.79      0.79     14717

[[2969  151   51  760]
 [  44 2491  281  427]
 [  17  340 1467    4]
 [ 436  552  100 4627]]

The Stacking Classifier is overfitting the training set and is not able to predict the test set perfectly as it has only 79% accuracy. Even though we have an improved accuracy by 1%, we reject this model due to overfitting!

This concludes that the best accuracy that can be obtained is 79% from Random Forest (With tuned Parameters). To further improve the accuracy of the model, we have to feed more rich data to the input.¶

	Unnamed: 0	525298014	17/03/04	0	719	7	904.34	0.1	1.1	...	646.29	613.33	32.96	.83	.00	.34	.28
0	NaN	925381920	17/07/22	1	51	25	8012.49	5	5	...	3934.73	3723.20	211.53	-7.20	6.04	-10.14	0.00
1	NaN	825437339	17/11/01	1	51	21	5977.00	0	3	...	3166.38	2882.41	283.97	-12.15	0.00	-4.30	0.00
2	NaN	689017953	17/10/14	1	51	22	5842.94	0	4	...	3051.38	3042.00	9.38	-12.80	0.00	-1.02	0.00
3	NaN	373396651	17/11/27	1	51	21	6018.78	0	2	...	3570.00	3570.00	0.00	-3.36	0.00	-0.61	0.00
4	NaN	588971563	17/06/19	1	51	26	7135.08	4	4	...	4102.56	4000.00	102.56	2.40	-1.92	-0.34	0.15

	cust_id	cust_status_ind	recency	frequency	monetary	NewVZPlan_plan_cnt	VerizonUNL_plan_cnt	billed_acct_size	VoiceTextData_pplangrp_cnt	VoiceData_revtype_cnt	...	gl_tot_service_rev_amt	gl_recur_rev_amt	gl_nonrecur_rev_amt	pplan_access_acrtn	tot_revenue_acrtn	access_rev_acrtn	eqp_rev_acrtn	usage_rev_acrtn	data_overage_amt	voice_overage_amt
count	7.581960e+05	758196.000000	758196.000000	758196.000000	758196.000000	758196.000000	758196.000000	758196.000000	758196.000000	758196.000000	...	758196.000000	758196.000000	758196.000000	758196.000000	758196.000000	758196.000000	758196.000000	758196.000000	758196.000000	758196.000000
mean	6.105298e+08	0.699624	197.979358	20.469335	3598.274701	0.785815	0.987718	2.305512	2.323062	2.151357	...	2293.113338	2194.204795	98.908543	9.599218	-13.440882	-2.575634	-2.685904	0.393704	1.164720	0.403131
std	2.304322e+08	0.458422	251.255878	8.726280	2740.704143	1.143805	1.751314	1.697620	1.752230	1.577893	...	1595.544941	1543.570243	217.113144	26.270767	40.332941	8.270055	20.061945	14.348179	10.694771	8.911295
min	2.252509e+08	0.000000	51.000000	1.000000	-10728.090000	0.000000	0.000000	0.000000	0.000000	0.000000	...	-7540.980000	-5443.390000	-9145.000000	-120.000000	-5070.900000	-355.340000	-5874.740000	-1315.370000	-0.040000	0.000000
25%	4.252560e+08	0.000000	51.000000	17.000000	1665.400000	0.000000	0.000000	1.000000	1.000000	1.000000	...	1200.707500	1127.955000	5.000000	0.000000	-13.020000	-2.410000	-1.870000	0.000000	0.000000	0.000000
50%	6.252509e+08	1.000000	51.000000	23.000000	3079.480000	0.000000	0.000000	2.000000	2.000000	2.000000	...	2072.735000	1985.830000	30.990000	2.260000	-4.710000	0.000000	-0.030000	0.000000	0.000000	0.000000
75%	7.890506e+08	1.000000	293.000000	27.000000	4999.102500	1.000000	1.000000	3.000000	3.000000	3.000000	...	3197.330000	3070.000000	119.500000	8.570000	-1.230000	0.000000	0.000000	0.000000	0.000000	0.000000
max	9.890513e+08	1.000000	962.000000	74.000000	95900.620000	30.000000	24.000000	22.000000	30.000000	28.000000	...	29032.510000	24626.960000	27599.570000	1893.330000	5969.420000	370.000000	2156.840000	4728.000000	2115.000000	4020.340000

Number of variables	53
Number of observations	735831
Total Missing (%)	0.0%
Total size in memory	297.5 MiB
Average record size in memory	424.0 B

Numeric	47
Categorical	0
Boolean	1
Date	0
Text (Unique)	0
Rejected	5
Unsupported	0

Distinct count	12
Unique (%)	0.0%
Missing (%)	0.0%
Missing (n)	0
Infinite (%)	0.0%
Infinite (n)	0

Minimum	0
5-th percentile	0
Q1	0
Median	0
Q3	0
95-th percentile	2
Maximum	11
Range	11
Interquartile range	0

Standard deviation	0.82867
Coef of variation	2.2742
Kurtosis	9.8599
Mean	0.36437
MAD	0.56934
Skewness	2.8866
Sum	268115
Variance	0.68669
Memory size	5.6 MiB

Value	Count	Frequency (%)
0	574875	78.1%
1	94054	12.8%
2	41538	5.6%
3	14285	1.9%
4	8184	1.1%
5	2229	0.3%
6	503	0.1%
7	107	0.0%
8	33	0.0%
9	14	0.0%
Other values (2)	9	0.0%

Standard deviation	0.96866
Coef of variation	1.4383
Kurtosis	5.5411
Mean	0.67345
MAD	0.7599
Skewness	1.8993
Sum	495549
Variance	0.93831
Memory size	5.6 MiB

Standard deviation	0.2708
Coef of variation	4.5979
Kurtosis	55.757
Mean	0.058897
MAD	0.11169
Skewness	5.9257
Sum	43338
Variance	0.073334
Memory size	5.6 MiB

Standard deviation	0.98385
Coef of variation	3.051
Kurtosis	16.999
Mean	0.32247
MAD	0.55668
Skewness	3.8353
Sum	237285
Variance	0.96796
Memory size	5.6 MiB

Standard deviation	1.5761
Coef of variation	1.4169
Kurtosis	3.6899
Mean	1.1124
MAD	1.1894
Skewness	1.7942
Sum	818526
Variance	2.4842
Memory size	5.6 MiB

Standard deviation	1.3845
Coef of variation	1.3832
Kurtosis	3.9193
Mean	1.0009
MAD	1.0153
Skewness	1.7908
Sum	736488
Variance	1.9167
Memory size	5.6 MiB

Standard deviation	0.67594
Coef of variation	1.87
Kurtosis	11.918
Mean	0.36147
MAD	0.52104
Skewness	2.5836
Sum	265982
Variance	0.45689
Memory size	5.6 MiB

Standard deviation	1.2845
Coef of variation	2.0241
Kurtosis	7.044
Mean	0.6346
MAD	0.91163
Skewness	2.516
Sum	466955
Variance	1.65
Memory size	5.6 MiB

Value	Count	Frequency (%)
0	415139	56.4%
1	205407	27.9%
2	76584	10.4%
3	23852	3.2%
4	10626	1.4%
5	3018	0.4%
6	842	0.1%
7	241	0.0%
8	73	0.0%
9	25	0.0%
Other values (7)	24	0.0%

Value	Count	Frequency (%)
0	415139	56.4%
1	205407	27.9%
2	76584	10.4%
3	23852	3.2%
4	10626	1.4%

Value	Count	Frequency (%)
0	697716	94.8%
1	33772	4.6%
2	3750	0.5%
3	407	0.1%
4	130	0.0%

Value	Count	Frequency (%)
0	635124	86.3%
1	37637	5.1%
2	27093	3.7%
3	15185	2.1%
4	10974	1.5%
5	5633	0.8%
6	2510	0.3%
7	1032	0.1%
8	384	0.1%
9	156	0.0%
Other values (6)	103	0.0%

Standard deviation	1.4114
Coef of variation	1.8782
Kurtosis	5.1648
Mean	0.75149
MAD	1.0446
Skewness	2.2066
Sum	552967
Variance	1.9921
Memory size	5.6 MiB

Standard deviation	1.3956
Coef of variation	2.0288
Kurtosis	6.6777
Mean	0.68787
MAD	0.99079
Skewness	2.4751
Sum	506158
Variance	1.9476
Memory size	5.6 MiB

Standard deviation	1.2238
Coef of variation	1.015
Kurtosis	2.8246
Mean	1.2057
MAD	0.91866
Skewness	1.4101
Sum	887159
Variance	1.4977
Memory size	5.6 MiB

Standard deviation	1.3306
Coef of variation	0.84533
Kurtosis	1.8558
Mean	1.574
MAD	1.0456
Skewness	1.213
Sum	1158227
Variance	1.7705
Memory size	5.6 MiB

Standard deviation	0.98021
Coef of variation	2.4406
Kurtosis	13.174
Mean	0.40163
MAD	0.63809
Skewness	3.2519
Sum	295529
Variance	0.96082
Memory size	5.6 MiB

Standard deviation	1.1824
Coef of variation	0.82195
Kurtosis	4.1995
Mean	1.4385
MAD	0.89743
Skewness	1.5183
Sum	1058487
Variance	1.398
Memory size	5.6 MiB

Standard deviation	1.0067
Coef of variation	2.4759
Kurtosis	13.114
Mean	0.40659
MAD	0.65024
Skewness	3.2615
Sum	299181
Variance	1.0134
Memory size	5.6 MiB

Standard deviation	1.1239
Coef of variation	1.4234
Kurtosis	5.0266
Mean	0.78961
MAD	0.86789
Skewness	1.846
Sum	581016
Variance	1.2632
Memory size	5.6 MiB

Standard deviation	0.92301
Coef of variation	1.5059
Kurtosis	10.053
Mean	0.61292
MAD	0.7104
Skewness	2.3773
Sum	451004
Variance	0.85194
Memory size	5.6 MiB

Standard deviation	0.28311
Coef of variation	4.5166
Kurtosis	72.985
Mean	0.062683
MAD	0.11853
Skewness	6.2361
Sum	46124
Variance	0.080153
Memory size	5.6 MiB

Project: Predicting The Life time Value of Customers¶

Importing Libraries¶

Importing Dataset¶

Sample details of first customer in the Database:¶

There are no Null values in the dataset.¶

Where are our customers located?¶

Outlier Removal¶

Removing unwanted columns:¶

Pandas Profiling¶

Overview

Variables

Correlations

Sample

Correlation Check¶

We see that these are the highly correlated features, so we have to remove them:¶

Let us remove:¶

Dropping Zip codes, as machine treats zip code with higher value as greater class:¶

K Means Clustering:¶

Determining the Number of clusters using Elbow Curve:¶

Checking the Attributes of Each Cluster:¶

Visualizing the Clusters in 3 Dimensions:¶

Cluster 0:¶

Cluster 0:¶

Overview

Variables

Correlations

Sample

Cluster 0 - Silver Findings:

Cluster 0:¶

Cluster 1:¶

Overview

Variables

Correlations

Sample

Cluster 1 - Platinum Findings:

Cluster 2:¶

Overview

Variables

Correlations

Sample

Cluster 2 - Bronze Findings:

Cluster 3:¶

Overview

Variables

Correlations

Sample

Cluster 3 - Gold Findings:

Cluster Naming:¶

Classification Models¶

Distribution of Output Variable:¶

Classification for 2019 Customers (New customers):¶

Standard Scaler for Feature Scaling¶

Decision Tree Classifier¶

Random Forest Classifier¶

Finding the best parameters for Random forest:¶

Random Forest tuned for Hyper Parameters¶

XG Boost Classifier¶

Naive Bayes Classifier¶

Ada Boost Classifier¶

Confusion Matrix for the Best Model¶

Shap Feature Importance: (Model Explainability)¶

Improving the Model Accuracy: (Grid Search)¶

The Best Estimators Obtained for the model after Grid Search CV are:¶

Results of Best Estimator:¶

Stacking¶

This concludes that the best accuracy that can be obtained is 79% from Random Forest (With tuned Parameters). To further improve the accuracy of the model, we have to feed more rich data to the input.¶

Standard deviation	1.4724
Coef of variation	1.6303
Kurtosis	4.8645
Mean	0.90318
MAD	1.0936
Skewness	2.0471
Sum	664591
Variance	2.168
Memory size	5.6 MiB

Standard deviation	0.68796
Coef of variation	3.0291
Kurtosis	17.303
Mean	0.22711
MAD	0.39386
Skewness	3.8472
Sum	167117
Variance	0.47329
Memory size	5.6 MiB

Standard deviation	1.3301
Coef of variation	1.9929
Kurtosis	7.1497
Mean	0.66742
MAD	0.95329
Skewness	2.4742
Sum	491106
Variance	1.7692
Memory size	5.6 MiB

Standard deviation	0.71179
Coef of variation	2.9183
Kurtosis	18.507
Mean	0.24391
MAD	0.41713
Skewness	3.857
Sum	179475
Variance	0.50664
Memory size	5.6 MiB

Standard deviation	0.56849
Coef of variation	2.1622
Kurtosis	14.03
Mean	0.26292
MAD	0.41344
Skewness	2.8512
Sum	193467
Variance	0.32318
Memory size	5.6 MiB

Standard deviation	1.4702
Coef of variation	1.7328
Kurtosis	5.6163
Mean	0.84847
MAD	1.0718
Skewness	2.2134
Sum	624327
Variance	2.1615
Memory size	5.6 MiB

Standard deviation	0.27487
Coef of variation	4.6463
Kurtosis	73.018
Mean	0.059159
MAD	0.11223
Skewness	6.3147
Sum	43531
Variance	0.075552
Memory size	5.6 MiB

Standard deviation	0.42203
Coef of variation	3.219
Kurtosis	18.885
Mean	0.1311
MAD	0.23485
Skewness	3.8806
Sum	96471
Variance	0.17811
Memory size	5.6 MiB

Distinct count	6967
Unique (%)	0.9%
Missing (%)	0.0%
Missing (n)	0
Infinite (%)	0.0%
Infinite (n)	0

Minimum	-355.34
5-th percentile	-13.54
Q1	-2.38
Median	0
Q3	0
95-th percentile	1.75
Maximum	370
Range	725.34
Interquartile range	2.38